PERFORMANCE ANALYSIS OF DEEP LEARNING ARCHITECTURES IN THE DETECTION OF FINDINGS IN MAMMOGRAPHIC IMAGES

Breast cancer is one of the leading causes of mortality among women in Brazil, making early detection a critical factor for therapeutic effectiveness. In this context, the present study aimed to conduct a systematic and comparative evaluation of ten Convolutional Neural Network (CNN) architectures pretrained on ImageNet, applied to the binary classification of digital mammograms into benign and malignant findings. The research employed the public VinDr-Mammo dataset, comprising more than 20,000 anonymized images, ensuring reproducibility and transparency. The methodological pipeline included standardized preprocessing, data augmentation, and transfer learning with replacement of the final classification layer. The evaluated architectures included ResNeXt-50, ResNet-101, DenseNet-121, MobileNet v2, EfficientNet-B0, ResNet-50, EfficientNet-B4, ShuffleNet v2, VGG16, and SqueezeNet 1.1. Performance was assessed using standard metrics such as accuracy, precision, recall, specificity, F1-Score, and AUC, complemented by confusion matrices and ROC curves. Results demonstrated limited initial performance, with AUC values ranging from 0.59 to 0.61, yet sufficient to establish a reproducible baseline for future optimization through fine-tuning. ResNeXt-50 achieved the most balanced error profile between false positives and false negatives, appearing as the most promising candidate for clinical scenarios. It is concluded that, although the initial performance is limited, this study contributes to the field of artificial intelligence in healthcare by providing a robust comparison of different architectures and highlighting the importance of progressive methodologies to achieve clinical robustness and generalization.